This release doesn't contain any deliberately-breaking changes. However, it contains a very complex new feature and while all of esbuild's tests pass, I would not be surprised if an important edge case turns out to be broken. So I'm releasing this as a breaking change release to avoid causing any trouble. As usual, make sure to test your code when you upgrade.
-
Implement the JavaScript decorators proposal (#104)
With this release, esbuild now contains an implementation of the upcoming JavaScript decorators proposal. This is the same feature that shipped in TypeScript 5.0 and has been highly-requested on esbuild's issue tracker. You can read more about them in that blog post and in this other (now slightly outdated) extensive blog post here: https://2ality.com/2022/10/javascript-decorators.html. Here's a quick example:
const log = (fn, context) => function() {
console.log(`before ${context.name}`)
const it = fn.apply(this, arguments)
console.log(`after ${context.name}`)
return it
}
class Foo {
@log static foo() {
console.log('in foo')
}
}
// Logs "before foo", "in foo", "after foo"
Foo.foo()
Note that this feature is different than the existing "TypeScript experimental decorators" feature that esbuild already implements. It uses similar syntax but behaves very differently, and the two are not compatible (although it's sometimes possible to write decorators that work with both). TypeScript experimental decorators will still be supported by esbuild going forward as they have been around for a long time, are very widely used, and let you do certain things that are not possible with JavaScript decorators (such as decorating function parameters). By default esbuild will parse and transform JavaScript decorators, but you can tell esbuild to parse and transform TypeScript experimental decorators instead by setting "experimentalDecorators": true
in your tsconfig.json
file.
Probably at least half of the work for this feature went into creating a test suite that exercises many of the proposal's edge cases: https://github.com/evanw/decorator-tests. It has given me a reasonable level of confidence that esbuild's initial implementation is acceptable. However, I don't have access to a significant sample of real code that uses JavaScript decorators. If you're currently using JavaScript decorators in a real code base, please try out esbuild's implementation and let me know if anything seems off.
⚠️ WARNING ⚠️
This proposal has been in the works for a very long time (work began around 10 years ago in 2014) and it is finally getting close to becoming part of the JavaScript language. However, it's still a work in progress and isn't a part of JavaScript yet, so keep in mind that any code that uses JavaScript decorators may need to be updated as the feature continues to evolve. The decorators proposal is pretty close to its final form but it can and likely will undergo some small behavioral adjustments before it ends up becoming a part of the standard. If/when that happens, I will update esbuild's implementation to match the specification. I will not be supporting old versions of the specification.
-
Optimize the generated code for private methods
Previously when lowering private methods for old browsers, esbuild would generate one WeakSet
for each private method. This mirrors similar logic for generating one WeakSet
for each private field. Using a separate WeakMap
for private fields is necessary as their assignment can be observable:
let it
class Bar {
constructor() {
it = this
}
}
class Foo extends Bar {
#x = 1
#y = null.foo
static check() {
console.log(#x in it, #y in it)
}
}
try { new Foo } catch {}
Foo.check()
This prints true false
because this partially-initialized instance has #x
but not #y
. In other words, it's not true that all class instances will always have all of their private fields. However, the assignment of private methods to a class instance is not observable. In other words, it's true that all class instances will always have all of their private methods. This means esbuild can lower private methods into code where all methods share a single WeakSet
, which is smaller, faster, and uses less memory. Other JavaScript processing tools such as the TypeScript compiler already make this optimization. Here's what this change looks like:
// Original code
class Foo {
#x() { return this.#x() }
#y() { return this.#y() }
#z() { return this.#z() }
}
// Old output (--supported:class-private-method=false)
var _x, x_fn, _y, y_fn, _z, z_fn;
class Foo {
constructor() {
__privateAdd(this, _x);
__privateAdd(this, _y);
__privateAdd(this, _z);
}
}
_x = new WeakSet();
x_fn = function() {
return __privateMethod(this, _x, x_fn).call(this);
};
_y = new WeakSet();
y_fn = function() {
return __privateMethod(this, _y, y_fn).call(this);
};
_z = new WeakSet();
z_fn = function() {
return __privateMethod(this, _z, z_fn).call(this);
};
// New output (--supported:class-private-method=false)
var _Foo_instances, x_fn, y_fn, z_fn;
class Foo {
constructor() {
__privateAdd(this, _Foo_instances);
}
}
_Foo_instances = new WeakSet();
x_fn = function() {
return __privateMethod(this, _Foo_instances, x_fn).call(this);
};
y_fn = function() {
return __privateMethod(this, _Foo_instances, y_fn).call(this);
};
z_fn = function() {
return __privateMethod(this, _Foo_instances, z_fn).call(this);
};
-
Fix an obscure bug with lowering class members with computed property keys
When class members that use newer syntax features are transformed for older target environments, they sometimes need to be relocated. However, care must be taken to not reorder any side effects caused by computed property keys. For example, the following code must evaluate a()
then b()
then c()
:
class Foo {
[a()]() {}
[b()];
static { c() }
}
Previously esbuild did this by shifting the computed property key forward to the next spot in the evaluation order. Classes evaluate all computed keys first and then all static class elements, so if the last computed key needs to be shifted, esbuild previously inserted a static block at start of the class body, ensuring it came before all other static class elements:
var _a;
class Foo {
constructor() {
__publicField(this, _a);
}
static {
_a = b();
}
[a()]() {
}
static {
c();
}
}
However, this could cause esbuild to accidentally generate a syntax error if the computed property key contains code that isn't allowed in a static block, such as an await
expression. With this release, esbuild fixes this problem by shifting the computed property key backward to the previous spot in the evaluation order instead, which may push it into the extends
clause or even before the class itself:
// Original code
class Foo {
[a()]() {}
[await b()];
static { c() }
}
// Old output (with --supported:class-field=false)
var _a;
class Foo {
constructor() {
__publicField(this, _a);
}
static {
_a = await b();
}
[a()]() {
}
static {
c();
}
}
// New output (with --supported:class-field=false)
var _a, _b;
class Foo {
constructor() {
__publicField(this, _a);
}
[(_b = a(), _a = await b(), _b)]() {
}
static {
c();
}
}
-
Fix some --keep-names
edge cases
The NamedEvaluation
syntax-directed operation in the JavaScript specification gives certain anonymous expressions a name
property depending on where they are in the syntax tree. For example, the following initializers convey a name
value:
var foo = function() {}
var bar = class {}
console.log(foo.name, bar.name)
When you enable esbuild's --keep-names
setting, esbuild generates additional code to represent this NamedEvaluation
operation so that the value of the name
property persists even when the identifiers are renamed (e.g. due to minification).
However, I recently learned that esbuild's implementation of NamedEvaluation
is missing a few cases. Specifically esbuild was missing property definitions, class initializers, logical-assignment operators. These cases should now all be handled:
var obj = { foo: function() {} }
class Foo0 { foo = function() {} }
class Foo1 { static foo = function() {} }
class Foo2 { accessor foo = function() {} }
class Foo3 { static accessor foo = function() {} }
foo ||= function() {}
foo &&= function() {}
foo ??= function() {}